ActionRAG's ablation shows a generic re-scorer earns flat two points on any store - and only a check that reads the reconciled structure scales with it.
A benchmark prices a question by its reuse across future tasks, then caps the budget inside one interaction. The ledger with no memory is not the agent.